AIScale - A coarse grained reconfigurable CNN hardware accelerator

Abstract

© 2017 IEEE. In this paper we propose a novel CNN hardware accelerator, called AlScale, capable of accelerating convolutional, pooling, fully-connected and adding CNN layers. In contrast to most existing solutions, AIScale offers a complete solution to the full CNN acceleration. AIScale is designed as a coarse-grained reconfigurable architecture, which uses rapid, dynamic reconfiguration during the CNN layer processing. Furthermore, a novel algorithm for mapping computations to the available computing resources enables AIScale to achieve higher utilization ratios than some of the previously proposed solutions. Results of the experiments indicate that the AIScale architecture is 1.16 to 2.73 times faster and consumes from 25% to 45% less energy on DRAM data transfers than the previously proposed MIT's Eyeriss CNN accelerator while using an identical number of computing units and having almost identical on-chip RAM memory size.